Fig. 6: The observed joint influence of motor repertoire and de novo learning experience extends to a more complex sinusoidal task.

a Networks were trained to produce ellipses with up to four different geometries by generating cosine and sine waves representing x and y positions, respectively. Networks were given a hold signal that indicated movement initiation, a continuous sine wave target signal that indicated the amplitude of the sine wave, and a categorical ‘one-hot-encoded’ cosine wave target signal. b Networks were trained on repertoires with different numbers of movements (from one to four) to model de novo learning. Within each repertoire, the movements could consist of either different sine waves and a constant cosine wave (‘varied sine wave’), or different cosine waves and a constant sine wave (‘varied cosine waves’). Following de novo learning, networks were separately trained to counteract (1) a sine wave perturbation, (2) a cosine wave perturbation, and (3) a reassociation perturbation where they had to produce different movements given learned target cues. The schematic indicates our predictions of what perturbations would be easier to counteract based on the center-out results. c Cosine waves (x position), sine waves (y position), and resultant elliptical output produced by an example network trained to produce sine waves with various amplitudes after de novo learning and each of the perturbations. d Loss during de novo learning and adaptation to each of the perturbations. Line and shaded surfaces smoothed mean and 95% confidence interval across networks of different seeds (n = 10 random seeds). Dashed line, loss of 0.2 included for easier comparison between loss in different networks. e, f Same as panels (c, d) but for networks trained on repertoires with varied cosine waves. For the reassociation perturbation, 45% of networks trained on either three- or four-movement repertoires with varied sine waves were able to adapt (MSE <0.4), compared to 70% for those with varied cosine waves. g Decay constants for exponential curves fitted to the loss curves in panels (d, f) for the sine and cosine wave perturbations.